10:45 AM - 12:15 PM
[PPS07-P24] Radio interferometric imaging of protoplanetary disks based on machine learning
Keywords:Radio interferometric imaging, Protoplanetary disk, Machine learning
A protoplanetary disk is a rotating disk composed of gas and dust, and is the site of planet formation. It is important to understand the general structure of the disk in order to theoretically understand planet formation, which requires information of the disk obtained by observation. In recent years, high-resolution observations of disks have been done by using large radio interferometers, and the brightness distribution of disks has been reconstructed by removing interference fringes from the interferometry data. However, since the reconstruction cannot provide a true image of the source, it is essential to compare the images obtained by various methods.
The purpose of this study is to develop a new method for detecting the detailed structure of protoplanetary disks from radio interferometry data. Recent studies have shown that sparse modeling techniques provide several times better resolution than the conventional CLEAN method. In this study, we examine whether the recently developed machine learning-based estimation can achieve the same or better performance in disk image restoration. Specifically, we attempted to reproduce images of disk models with multiple gaps by training a neural network called U-Net to restore disk models with concentric gap structures. U-Net is one of the neural networks for image processing by machine learning, and is used for image segmentation because of its excellent ability to identify structures even over noise. If U-Net is applied to the observation of dust continuous waves by interferometry, it may be possible to improve angular resolution and image fidelity compared to conventional restoration methods such as CLEAN. To verify this, we compared disk images restored by U-Net and CLEAN with respect to the position, width, and depth of the gap structure.
The results show that U-Net accurately identifies the position and width of the gap structure with at least twice the resolution of CLEAN for disk models that are close to the training data. For disks with structures different from the training data, U-Net is able to estimate the brightness distribution with higher resolution than CLEAN. On the other hand, CLEAN outperforms U-Net in restoring the depth of the gaps. It is also found that U-Net sometimes misrecognizes structures that are not included in the training data and returns results that differ from the true distribution.
The results of this study indicate that U-Net may be able to recover finer concentric disk structures than conventional methods. On the other hand, the method needs to be improved to reliably recover non-axisymmetric structures such as spirals and vortices.
The purpose of this study is to develop a new method for detecting the detailed structure of protoplanetary disks from radio interferometry data. Recent studies have shown that sparse modeling techniques provide several times better resolution than the conventional CLEAN method. In this study, we examine whether the recently developed machine learning-based estimation can achieve the same or better performance in disk image restoration. Specifically, we attempted to reproduce images of disk models with multiple gaps by training a neural network called U-Net to restore disk models with concentric gap structures. U-Net is one of the neural networks for image processing by machine learning, and is used for image segmentation because of its excellent ability to identify structures even over noise. If U-Net is applied to the observation of dust continuous waves by interferometry, it may be possible to improve angular resolution and image fidelity compared to conventional restoration methods such as CLEAN. To verify this, we compared disk images restored by U-Net and CLEAN with respect to the position, width, and depth of the gap structure.
The results show that U-Net accurately identifies the position and width of the gap structure with at least twice the resolution of CLEAN for disk models that are close to the training data. For disks with structures different from the training data, U-Net is able to estimate the brightness distribution with higher resolution than CLEAN. On the other hand, CLEAN outperforms U-Net in restoring the depth of the gaps. It is also found that U-Net sometimes misrecognizes structures that are not included in the training data and returns results that differ from the true distribution.
The results of this study indicate that U-Net may be able to recover finer concentric disk structures than conventional methods. On the other hand, the method needs to be improved to reliably recover non-axisymmetric structures such as spirals and vortices.